← Back to Blog

guide

By ReclaimLLM Team

How to estimate Cursor cost per session

Estimate Cursor session cost from request-level usage, account for cache reads and writes, and understand what a team spend limit covers.

A Cursor spend limit tells you how much a team or member can spend under the applicable billing setting. To explain why one coding session used more than another, you need request-level usage. A single chat can involve multiple model calls, and each call can bill differently depending on the model, output, and how much of the prompt the provider served from cache.

Start with the conversation ID and group the model requests that belong to it. Price each request using its reported usage and the rates that apply to that model and plan. Label the result as an estimate and reconcile aggregate totals against the billing dashboard. Cursor spend management · Cursor staff explanation of cost telemetry

What a Cursor spend limit covers

Cursor distinguishes spending alerts from limits. Alerts tell someone that usage crossed a threshold. Use the applicable team spend limit or additional-usage setting if the goal is to cap spending. Confirm the setting after creating a team, especially if developers previously managed spending under individual plans. The scope and dashboard location depend on the plan. Cursor spend management · Team setup discussion

Session attribution answers a different question. If a team approaches its limit, which sessions drove it? If a long agent run looks expensive, did it generate more output, switch models, or repeatedly process a large prompt at the uncached rate? The same team total can come from very different workflows.

Build a session estimate from model requests

Cursor staff describe Enterprise OpenTelemetry Export as a source of per-request events with token counts, model, user, and a conversation ID. Staff also describe desktop hooks that provide token counts after a parent turn finishes. Those hook payloads provide a narrower view and contain no synchronous cost figure at prompt submission. Cursor staff explanation

  1. Group by conversation. Keep the request timestamp, surface, and model. Include child requests where the telemetry associates them with the conversation. Track requests with missing IDs separately so they stay visible.
  2. Normalize token categories. Identify uncached input, cache reads, cache writes where separately priced, and output. Some APIs report cached input as a subset of input, while others expose separate categories. Charge a token once. Anthropic prompt caching · OpenAI prompt caching
  3. Apply rates per request. Use the applicable model, variant, rate date, unit, and plan terms. For separately priced categories, multiply the token count by its rate per million tokens and divide by one million. Sum those line items across the session. Track external tool charges separately if they appear on a different invoice.
  4. Reconcile totals. Compare totals for the same account and billing period with dashboard usage. Document missing events, included credits, and any difference between the rates used in your estimate and the rates on the invoice.

Subagents can use a different model from the parent, so pricing the whole session at the parent model's rate can skew the result. See how to audit Cursor subagent models.

Check for separate causes before blaming a cache miss. In one case, staff traced an unexpected charge to an older chat that started a new turn when it reopened. In another, an agent submitted the same paid external job several times. Record the trigger for each model request when available, and match paid tool calls to their job IDs and provider charges. Reopened chat report · Duplicate paid job report

Worked example with illustrative rates

Suppose two calls each process a 100,000-token input and generate 5,000 output tokens. Across the two calls, 110,000 input tokens are uncached and 90,000 are cache reads. Assume no separately priced cache writes and use these illustrative rates:

CategoryTokensExample rate per millionEstimated amount
Uncached input110,000$3.00$0.330
Cache reads90,000$0.30$0.027
Output10,000$15.00$0.150
Total$0.507

If the second call's 90,000 cache-read tokens had instead billed as uncached input, the estimate would be $0.750. The $0.243 difference isolates the effect of the cache in this example. Actual charges depend on the model and plan, including any cache-write rate or billing adjustments. These figures illustrate the calculation using synthetic data.

Large context makes cache misses easier to see

In a Cursor support case involving Gemini Carbon, staff confirmed that seven requests reprocessed the full context at the uncached input rate within a half-hour window. The context was roughly 280,000 to 490,000 tokens. Staff observed that the provider cache appeared to last only a minute or two at that size in that run. Some misses followed long-running responses within an agent turn. The case gives you a pattern to investigate in a usage trace, with an observed duration specific to that workload. Cursor Gemini cache report

When you review an expensive session, compare cache reads and uncached input across adjacent calls on the same model. Check the request timing as well as the visible time between chat messages. Elapsed time can explain some misses. A different prefix, model, or cache policy can also change which input the provider can reuse. Anthropic documents a five-minute default for Claude API prompt caching and an optional one-hour duration. OpenAI documents separate retention behavior depending on the model and configuration. Use the provider's documentation and your own trace together. Anthropic prompt caching · OpenAI prompt caching

For cache mechanics and measured token counts, see how we built token savings for coding agents and how to measure token waste in coding sessions.

Where Claude Code cost tracking fits

For Claude Code, start with the usage commands and plan-specific reporting Anthropic documents. Teams that collect telemetry can inspect cache and model usage using the relevant organization or cloud billing records. Track subscription usage separately from metered API charges when comparing sessions across tools. Claude Code cost documentation

FAQ

Will a Cursor spend alert stop a session?

An alert sends a notification. Use the applicable spend controls to cap spend and confirm how the limit behaves on your team's plan. Cursor spend management

Can a Cursor hook see the cost before a prompt runs?

Cursor staff describe hooks as receiving no synchronous cost field. Some desktop hooks report parent-turn tokens after the turn completes. A request ledger built from telemetry can estimate spend already incurred, with a delay that depends on when events arrive. Cursor staff explanation

What should I compare when a session suddenly costs more?

Check the executed model and variant, the output volume, and the split between uncached input, cache reads, and cache writes. Compare request-by-request to find which change explains the difference.